基于仿生技术和反向传播神经网络的黄芪产地判别模型构建研究
Identification of origin place for Astragali Radix based on biomimetics
陈万金 1李虹 1张沛沛 2邵炜娴 1王越 1范昕煜 1赵婷 3刘凤波 3魏胜利 3于芳 1张媛3
作者信息
- 1. 北京中医药大学中药学院,北京 102488
- 2. 中国药科大学药学院,南京 211198
- 3. 北京中医药大学中药学院,北京 102488;中药材规范化生产教育部工程研究中心,北京 102488
- 折叠
摘要
目的 基于仿生技术和反向传播神经网络(BPNN)构建黄芪的产地鉴别模型.方法 采用色度计、电子鼻和电子舌共测得21项指标,通过RFI进行特征筛选后得到14项指标,并将黄芪产地鉴定问题建模为多分类问题.通过对随机森林(RF)、支持向量机(SVM)和BPNN这三种机器学习模型的比较,我们建立了一个基于BPNN的黄芪产地分类决策系统.结果 BPNN仅用了 11个特征变量就能够较好地预测黄芪产地.多分类模型构建后,引入SHAP值对构建的产地鉴别模型进行解释.结论 SHAP特征重要性的排序揭示了变量在实际构建出的模型的重要程度.可解释预测模型在增加产地预测模型的透明度的同时,又能保持原模型的判别正确率.该研究为产地鉴别模型的构建提供了一定的启示,也为客观产地鉴别提供了参考.
Abstract
Objective To construct the origin identification model of the roots of Astragalus membranaceus var.mongholicus based on biomimetics and back propagation neural network(BPNN).Methods Totally 21 indicators were measured by colorimeter,electronic nose(E-nose),and electronic tongue(E-tongue).Totally 14 indicators were obtained by random forest importance(RFI)after feature screening,and AR origin identification was modeled as a multi-classification problem.By comparing the three machine learning models(RF,SVM,and BPNN),a decision system was built for classification based on BPNN.Results BPNN well predicted the geographical origins of AR with only 11 feature variables.After constructing the multi-classification model,SHapley Additive exPlanation(SHAP)values were introduced to interpret the constructed origin identification model.Conclusion The importance ranking of the SHAP features shows how important the variables are in the actual model.Interpretable prediction models increase the transparency of the origin prediction model while maintaining the discrimination correctness of the original model.This study provides some reference for the construction of origin identification models.
关键词
黄芪/产地鉴别/仿生技术/反向传播神经网络/SHAP/可解释机器学习Key words
Astragali Radix/geographical traceability/biomimetics/back propagation neural network/SHapley Additive exPlanation/interpretable machine learning引用本文复制引用
出版年
2024
NETL
NSTL
万方数据